Tracking Progression from Mild Cognitive Impairment to Alzheimer's Disease Using Multivariate Biomarkers and Pattern Classification
Wei Wen,Yue Cui,Bing Liu,Xiantong Zhen,Ming Fan,Suhuai Luo,Tao Liu,Wanlin Zhu,Mira Park,Tianzi Jiang,Jesse Jin
DOI: https://doi.org/10.1016/j.jalz.2011.05.354
2011-01-01
Abstract:Prediction of conversion from mild cognitive impairment (MCI) to Alzheimer's disease (AD) has attracted significant attention recently. Most investigations have used a single modality of data. In the present study, multiple features from different modalities of data, including structural MRimaging (MRI), cognitive measures (CM) and cerebrospinal fluid (CSF) biomarkers, were examined in the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, in order to establish a set of sensitive predictors and to capture MCI conversion and corresponding conversion time. Based on their follow-up evaluations, MCI subjects at baseline were categorized into two subgroups: MCI converters (MCIc), MCI individuals who developed AD within the 24 months follow-up period, and MCI non-converters (MCInc), MCI individuals who did not developed AD within 24 months. The prediction procedure consisted of three stages: (1) feature extraction and collection; (2) feature selection; and (3) classification. MRI features which had been extracted automatically using FreeSurfer, as well as a set of typical cognitive measures and CSF biomarkers were downloaded from ADNI website. Then mutual information-based feature ranking followed by a wrapper method was employed to select an optimal subset of MRI features and cognitive features. After feature selection, support vector machines were trained using both a single modality and combinations of multiple modalities of features. Note that training was conducted on NC and AD individuals and testing was conducted on MCI individuals. A real predictive value was generated for each test subject using the classifier, which indicated whether the subject was classified as MCIc or MCInc. Selected MRI features were measures of the entorhinal cortex, hippocampus, middle temporal gyrus, inferior parietal cortex and retrosplenial cortex. Selected cognitive assessments included CDR-SOB (Clinical Dementia Rating Scale Sum of Boxes scores), logical memory delayed recall, MMSE (MiniMental State Examination), number cancellation, ADAS (Alzheimer's Disease Assessment Scale-Cognitive subscale) word list delayed recall, FAQ (Functional Assessment Questionnaire), logical memory immediate recall, AVLT (Auditory Verbal Learning Test) delayed recall, clock drawing and ADAS word list immediate recall. The best performance was attained using multiple features, including select cognitive, MRI features and t-tau/Aß 1-42 (ratio of total tau to amyloid-ß 1 to 42 peptide), which achieved 96.43% sensitivity, 35.63% specificity, 59.44% accuracy and an AUC (area under a receiver operating characteristic curve) of 0.7588 by 24months follow-up. The specificity rose from 22.63% at 6 months to 35.63% at 24months. The sensitivities were always above 96% from 6 to 24 months. As a result, the accuracy increased from 25.87% to 59.44%. Predictive values of MCI subjects who converted within 1 year and after 1 year were -0.65±0.26 and-0.39±0.28, respectively, which are significantly different (p > 0.001). Our selected features, obtained using automatic feature selection techniques, correspond with the findings of recent studies. The combination of multiple features including CM, MRI and CSF biomarkers, out performed the use of a single modality of features because the different features selected were mutually complementary. In addition, our classifier was able to predict 1-year conversion time. Taken together, this study establishes multivariate predictive markers which may be practical and helpful for clinical diagnosis.